Research On Clustering Methods For Analyzing Overlapping Local Gene Expression Patterns

Overlapping local gene expression patterns means the expression characters of agroup of genes determined by underlying biological mechanisms of genes expressed onlyunder a specific sub set of conditions, and genes expressed under diferent sub set ofconditions. Research on mining and analyzing such patterns can be helpful for revealinggene function involved in diferent cellular regulation phases and biological relationshipsbetween diferent gene groups. However, due to the specific demand and difculty ofoverlapping local gene expression pattern analysis which bring forward great challengefor research of clustering methods, existing clustering methods can not fully fulfill thesedemands, which calls for novel and efective clustering methods.In this dissertation, we focus on the design and application of new clustering meth-ods to solve four related problems in overlapping local gene expression pattern analysis,i.e., mining local gene expression patterns with overlapping genes between gene cluster-s, mining local gene expression patterns with overlapping genes and subspaces betweengene clusters, determining the subspace boundary of local gene expression patterns, andextracting and optimizing local gene expression patterns. Based on the fundamentalsof both fuzzy clustering and variable-weighting-based subspace clustering methods, theproposed methods include fuzzy hard subspace clustering method, fuzzy soft subspaceclustering method, Changing Window method, and Post-processing method.The main research roadmap of this dissertation includes tow phases. On one hand,on the basis of review existing clustering methods, we proposed tow clustering methods,i.e., fuzzy W-K-Means and fuzzy EWKM, which combine the advantages of fuzzy clus-tering and variable-weighting-based clustering. The two proposed clustering methods cannot only find diferent group of genes which show high expression similarity under a subset of conditions, but also map individual genes on multiple clusters. Experiment result-s have proven that the resulting locally expressed gene clusters show high consistencemeasured by GO-based cluster validation method. On the other hand, we proposed theChanging Window method and Post-processing method to determine the subspace bound-ary of each locally expressed gene cluster and to reduce the disturbance caused by noisegenes in individual gene clusters. The Changing Window method provides an interactive visual analysis environment under which various local gene expression patterns under dif-ferent subspaces can be extracted. Moreover, the Post-processing method provides severalsteps including the identification and elimination of noise genes, the remodeling of localgene expression patterns, and the reuse of the refined local patterns, so as to improve thesignificance of the extracted local gene expression patterns.Although individual clustering analysis methods proposed in this dissertation focuson diferent aspects of the problems existing in local gene expression pattern analysis,the proposed clustering analysis methods can be used in a supplementary way to serveas an integral solution to the problems of the mining and extraction of local patterns, theoptimization local patterns, and the reuse of the refined patterns.